Tafla, calculations you can understand and trust

Comments

0

Hörður RagnarssonMaker

What features would help Tafla explain calculations even better?

About Tafla, calculations you can understand and trust

I built Tafla so you can ask a math question and have the answer explained on a whiteboard instead of in a chat, similar to how a teacher would walk you through a problem in a classroom. An AI model sets up the calculation, a calculation engine does the actual math, and Tafla narrates the result step by step. You can see every formula, change any number and watch the whole thing recalculate.

I use Claude Code a lot, and often it has already done the math by the time I want to see it laid out. So Tafla now works from there too. When Claude has worked out a number, ask it to show you in Tafla. It hands the calculation over, Tafla draws it up on the whiteboard, and Claude gives you a link. Because Tafla's engine recalculates everything from scratch, Claude also gets to check its own arithmetic against it. The link works for anyone, no account needed to open it.

It's free to use. To connect it, add https://www.tafla.is/api/mcp as an MCP server in Claude Code (it works in Claude.ai, Codex and Cursor too), sign in with your Tafla account once, and you're set. The details are at https://www.tafla.is/agents. I'd love to see what you make with it.

How it was built with Claude

I built Tafla using Claude Code and learned a lot along the way. For example, by default, LLMs create bad captions. Just imagine subtitles filled with em dashes and colons.

When designing, I often asked it to generate HTML artifacts with a couple of different ideas I had, which made prototyping incredibly fast. At one point Claude started using a Claude Design like canvas instead, which was slower and more expensive, so I added a rule to my global instructions telling it to stick to HTML artifacts. Cannot recommend this more.

Claude Code has also been a great partner for benchmarking different models to figure out which ones work best in terms of speed, accuracy and cost. I ended up building a benchmark comparing LLMs doing their own arithmetic, using a code sandbox, or using Tafla's calculation engine. So far I have run over 500 tests across 19 problems. One interesting finding was that cheaper models can often get you results faster without sacrificing accuracy, since the engine handles the actual calculations.